💬 NLP

Beyond Isolated Behaviors: Hierarchical User Modeling for LLM Personalization

This paper introduces PHF, a sociologically grounded framework inspired by Bourdieu's Theory of Practice that enhances LLM personalization by modeling user data through a hierarchical structure of practices, habitus, and fields, achieving consistent performance improvements on the LaMP benchmark.

Original authors: Liang Wang, Xinyi Mou, Xiaoyou Liu, Tiannan Wang, Yuqing Wang, Zhongyu Wei

Published 2026-06-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Liang Wang, Xinyi Mou, Xiaoyou Liu, Tiannan Wang, Yuqing Wang, Zhongyu Wei

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to get a very smart, but slightly generic, AI assistant to understand you specifically. Right now, most AI personalization works like a scrapbook of isolated moments. If you liked a Harry Potter book last year and an Avengers movie last month, the AI just sees two separate facts: "User liked Harry Potter" and "User liked Avengers." It doesn't really understand the story of your taste or how your interests have evolved.

This paper proposes a new way to think about personalization, called PHF (Practice-Habitus-Field). It borrows ideas from a famous sociologist named Pierre Bourdieu to organize user behavior into three layers, like a set of Russian nesting dolls.

Here is how the three layers work, using simple analogies:

1. Practice (The Single Action)

The Analogy: Think of this as a single footprint in the sand.

What it is: Every time you click a link, rate a movie, or ask a question, that's a "practice."
The Problem: Footprints get washed away by the tide (noise) or look different depending on the angle. If the AI just looks at one footprint, it might think you like "sand" because you walked on the beach once, even if you actually hate the beach.
The Fix: The paper's system, PHFCompass, acts like a noise-canceling filter. It looks at your single action and asks, "What is the essence of this?" It turns a messy, specific action into a clean, clear "semantic ID" (like a barcode) that represents the true intent, ignoring the surface-level clutter.

2. Habitus (Your Personal Style)

The Analogy: Think of this as your muscle memory or your personality.

What it is: If you leave footprints every day for a year, they form a path. That path is your "Habitus." It's not just a list of what you did; it's your stable tendency. It captures that you used to love fantasy books but have slowly shifted toward superhero stories.
How it works: The system takes all those cleaned-up "practices" and blends them together over time. It gives more weight to what you did recently (because your taste changes) but keeps the long-term pattern. This creates a "User Profile" that feels like a real person with a history, not just a list of data points.

3. Field (The Tribe)

The Analogy: Think of this as joining a club or a neighborhood.

What it is: Even if you are unique, you share habits with others. If you and 100 other people all have a "Habitus" that loves sci-fi and hates horror, you belong to the same "Field."
Why it helps: Sometimes, you don't have enough history for the AI to understand you (maybe you are new or very quiet). But if the AI knows you belong to the "Sci-Fi Field," it can borrow insights from everyone else in that club to guess what you might like. It's like saying, "I don't know your specific taste yet, but since you hang out with this group, here is what they usually enjoy."

How It All Fits Together

The paper introduces a tool called PHFCompass to build this system. It doesn't require retraining the giant AI model (which is expensive and slow). Instead, it acts like a smart translator that sits in front of the AI.

It takes your raw history and turns it into clean Practices.
It blends those into your personal Habitus.
It finds your Field (your tribe) based on who else is similar to you.
It feeds all three layers to the AI at once.

The Result:
When you ask the AI a question, it doesn't just see "User asked about X." It sees:

"This is a user who has a long-term habit of liking X (Habitus)."
"This user is part of a group that generally prefers Y (Field)."
"Here is the specific, noise-free intent of this question (Practice)."

What the Paper Found

The authors tested this on a benchmark called LaMP (which involves tasks like writing news headlines, tagging movies, or rating products).

The Outcome: Their method (PHFCompass) consistently beat the old methods.
Why: The old methods were like looking at a pile of bricks (flat data). PHFCompass built a house (hierarchical structure). It showed that understanding how behaviors connect over time (Habitus) and who you are similar to (Field) makes the AI much better at guessing what you want.

In short, the paper argues that to truly personalize an AI, we need to stop treating users as a list of random clicks and start treating them as people with a history, a style, and a community.

Technical Summary: Beyond Isolated Behaviors: Hierarchical User Modeling for LLM Personalization

Problem Statement

Despite the remarkable capabilities of Large Language Models (LLMs) across diverse domains, personalizing their outputs to individual users remains a significant challenge. Existing approaches predominantly adopt a flat behavioral paradigm, treating user interactions as an unordered collection of isolated instances. This formulation leads to two critical limitations:

Behavior Fragmentation: Individual user histories lack internal structure, failing to consolidate isolated behaviors into stable, long-term patterns.
User Isolation: Users are modeled independently, ignoring shared behavioral regularities among users with similar profiles. This limits generalization, particularly in scenarios with sparse user history.

The paper argues that effective personalization requires a structured view that models both how behaviors consolidate within an individual and how regularities are shared across a user population.

Methodology: The PHF Framework

Drawing on Pierre Bourdieu's Theory of Practice, the authors propose PHF (Practice-Habitus-Field), a sociologically grounded, hierarchical framework for LLM personalization. This framework conceptualizes user behavior across three levels:

Practice: Individual, episodic behaviors (e.g., a single query-response pair).
Habitus: The temporal accumulation of practices into stable, long-term behavioral dispositions unique to a user.
Field: Shared behavioral regularities and collective patterns among users with similar dispositions.

PHFCompass Implementation

To instantiate PHF, the authors developed PHFCompass, a lightweight, model-agnostic architecture that operates on a frozen LLM. The system functions through four key stages:

Denoised Practice Extraction (Semantic Quantization):
Raw user behaviors are encoded and subjected to multi-level residual vector quantization. This process abstracts continuous interactions into discrete Semantic IDs (prototypes). By mapping behaviors with different surface forms but similar intents to shared prototypes, the system filters surface-level noise while preserving core behavioral semantics.
Habitus Construction (Temporal Aggregation):
The sequence of denoised practice representations for a user is aggregated over time using a Time Fusioner. To account for the recency effect without letting recent behaviors dominate, the model employs a temporally weighted aggregation scheme using square-root scaling. This results in a Habitus embedding that encodes the user's stable, long-term behavioral disposition.
Field Formation (Collaborative Clustering):
Users are organized into latent Fields based on their Habitus embeddings. Using K-Means clustering, users with similar long-term dispositions are grouped. Each cluster centroid serves as a Field embedding, capturing collective behavioral regularities. This allows users to benefit from shared signals, particularly aiding in cold-start or sparse-history scenarios.
Hierarchical Generation:
During inference, the LLM is conditioned on a unified input sequence comprising the Habitus embedding (individual disposition), the Field embedding (collective pattern), and the current Query embedding. Only lightweight projection layers are trained, leaving the underlying LLM frozen.

Key Contributions

Theoretical Framework: The proposal of PHF, a unified hierarchical model that integrates individual behavioral formation (Practice $\to$ Habitus) with cross-user structural regularities (Field), grounded in sociological theory.
PHFCompass Architecture: A concrete, lightweight implementation that operationalizes these concepts via residual behavioral abstraction, temporal aggregation, and collaborative clustering without requiring modifications to the base LLM.
Empirical Validation: Extensive experiments demonstrating that modeling behavioral structure hierarchically outperforms flat aggregation and retrieval-based baselines.

Experimental Results

The authors evaluated PHFCompass on the Language Model Personalization (LaMP) benchmark, covering six tasks (three classification, three generation).

Performance: PHFCompass achieved consistent state-of-the-art improvements across all tasks compared to strong baselines, including retrieval-based methods (BM25, Dense), aggregation-based methods (OPPU, PPlug), and zero-shot models.
- In classification tasks (LaMP-1 to 3), the model showed substantial gains, attributed to the field component compensating for sparse individual histories.
- In generation tasks (LaMP-4, 5, 7), the model maintained competitive performance, benefiting from robust practice and habitus modeling.
Ablation Studies:
- Removing the Habitus component caused the largest performance degradation, confirming that fine-grained individual behavioral representation is fundamental.
- Removing the Field component led to moderate declines, particularly in classification, validating the utility of collaborative signals.
- Replacing temporal weighting with uniform averaging and removing residual quantization both resulted in performance drops, highlighting the importance of temporal sensitivity and behavioral abstraction.
Extensibility: The framework demonstrated robustness across different encoder-decoder combinations (e.g., Contriever/Qwen, RoBERTa/LLaMA), confirming its model-agnostic nature.
Interpretability: Analysis of Semantic IDs revealed that the hierarchical quantization successfully disentangles user-specific intent from shared semantic structures, with specific codes corresponding to coherent thematic groups (e.g., fantasy romance vs. identity-focused romance).

Significance and Claims

The paper claims that PHF represents a paradigm shift from treating user data as a flat collection of instances to modeling it as a structured, hierarchical system. By explicitly capturing the temporal coherence within individual behavior and the shared regularities across users, PHF addresses the fragmentation and isolation inherent in current personalization methods.

The authors emphasize that their approach provides a unified perspective that is both interpretable (through discrete semantic prototypes and cluster fields) and extensible (model-agnostic). While the current work is validated on the LaMP benchmark, the framework is presented as a generalizable approach for organizing user behavior in LLM personalization, offering a sociologically grounded alternative to purely statistical or retrieval-based methods.

The paper acknowledges limitations, noting that experiments are currently restricted to English text-based tasks with static histories, and that the specific implementations (e.g., K-Means for fields) are just one possible instantiation of the broader PHF concept.